|
PureBytes Links
Trading Reference Links
|
Bilo,
I take exception to your generalizations about NNs.
Your comments reflect popular consensus circa 1990. Much has been learned
since then about proper techniques that makes all the difference. Proper
data processing (detrending, filtering, etc.). Proper training, testing
and validation. And above all, proper representation of the data
(Euclidean space, frequency space, Wavelet space, embedded space, phase
space, etc.) When all the necessary techniques are applied, results can be
exceptional and reliable.
>>
in my experience from a few years back has been this: you can basically do
two things with them:
1. curve fit the crap out of non linear time series and then sit back and
ponder at great stats (90% profitable not uncommon )
and paper profits. try trading it in real time - better make a donation to
the crippled children fund.
<<
A result of poor modeling technique.
>>
2. prune it and it turns into an overgeneralizer you then are better off
just using rule based approach( aka go trade aberration ) it will do the
same thing.
<<
Some nonlinear models do not lend themselves to modeling financial time
series. No amount of pruning will make it any better.
>>
also a very common experience with nn is someone comes out with real good
nn trading results because they catch a streak or an up trend and anything
works during that time including nn. as soon as the trend ends.nn is a
toast. so they have to retrain and the cycle go on like that till they
finally ditch the thing.
<<
Proper performance analysis calls for validation on a sufficiently long
data series. I train on 4 years, test on another 4 years and then validate
on yet another 5 years. A good model will hold up throughout.
>>
all nn's are good at is what your brain is good at: look at the chart and
see how easy it is to pick tops and bottoms.
top looks like a top and bottom looks like a bottom. well same thing goes
about nn's. you can make them recognize
those with precision, but....only in 20/20 hindsight. nn are not better off
predicting the those in real time than your brain is... try picking tops
and bottoms in real time... same goes for nns.
<<
This is a false analogy. Although you and the NN are trying to pick tops
and bottoms, the human mind can process up to only 9 items of information
effectively when making a decision. Beyond that, the decision space is too
complex and large to cogitate. However, NNs can process higher dimensions
and larger sets of interacting rules much more readily. You just need to
know how to properly exploit that power, without over-optimizing. There
is, however, the well known "curse of high dimensionality" that affects all
models, but this is where a wise choice of state space representation comes
into play.
>>
as far as building price prediction models - price is mainly
unpredictable... it's too dynamic
<<
The goal here is not to predict the market, but to make wise trading
decisions. For that, you do not necessarily need to predict where price
will be. It's like playing poker: you don't know exactly what hand your
opponent has, but you can still formulate an optimal playing strategy.
>>
input number limitation ( over generalization vs curvefit ), input
selection. training process, black box ( rules are unknown ), etc. are the
soft spots.
<<
By "soft spots" I presume you mean "areas that require skill and
understanding".
>>
imho, basically it's a good tool if you want to paper trade forever. i
rejected nns and products like abtech as a primary model a long time ago.
best thing you can do with them is pattern recognition in hindsight...
<<
It's not fair to dump on the GMDH modeling process. It is a valid
approach. I'm willing to bet failure was due to inadequate data
preprocessing and representation.
>>
however i wholeheartedly endorse fuzzy logic concepts. too bad there are no
clean FL trading products out there. FL is an extremely valuable tool.
<<
Have you played with SAFIR-X by Jewelsoft? ( http://www.sirtrade.com )
>>
all i am seeing to this day is 90% paper profits on past data and marginal
performance on real data...
<<
This suggests the simulated trading and/or the validation processes were
not realistic enough.
There's more to successsful modeling than merely training a NN. I estimate
that 80% of the effort is in finding the right indicators. Remember:
garbage in --> garbage out. That leaves 10% for NN optimization and 10% for
testing and validation.
Mark Jurik
Jurik Research
http://www.jurikres.com
PS. I am not in the system vending business. My systems are for
proprietary use only.
|